Lightweight U-Net Image Dehazing Network for Remote Airport Tower Control
Qimin Ren, Wujun Xu · 2024
Image dehazing is a significant part of low-level vision tasks in the area of airport traffic surveillance. However, the existing image dehazing methods based on deep learning suffer from the limitation on both model generalization and computational cost. A lightweight U-Net dehazing network based on improved inverted residual block (IIRB) and adaptive detail enhancement (ADE) module is proposed for remote airport tower control. Firstly, the low and mid-frequency feature was extracted with IIRB module from image smooth regions. Secondly, the high- frequency details and textures were extracted with ADE module, which further led to the enhancement of the network's expression in sky regions and detail features. Finally, the model was trained combining L1 loss and contrastive loss, which aimed to improve the model's generalization. To evaluate the effectiveness and practicality of the proposed method, the experiments were conducted both on SOTS dataset and real samples of airport hazy images. The experimental results show that, compared to lightweight AOD and weakly supervised RefineD, the proposed method achieves a balance between visual quality and real-time performance.